Thank you for coming. It's really cool to see you here . And let's start with one number to get through this presentation. So, it's zero. And today I want to say why it's so important to talk about this .
Part one, we at Gearbox are crazy about security. So we have n't had any security incidents in the last 5 years, which is great. The second part great. The second part , we do not have any bad debt for this part. And of course, when we talk about how to manage capital with agents, we want to ensure all reliability, security, and protection.
Because, as you know, an agent can behave very inappropriately. And zero could be how much money will be left in your account if you don't provide proper agent protection. And my story today is about my personal experience, what I've learned while piloting various experiments at Gearbox over the past two years. We will start looking into this topic in 2024, building some prototypes on EigenLayer and learning how we do it right. In this story, I also want to talk about Gearbox's agent framework and how we see the best way to organize capital management capital management .
And finally, I can offer a really good way for you to set up an AI wall with your own storage right now. So, it is an autonomous way of managing your capital as a structured product without a curator. And first I want to talk about the numbers, because we all need to understand together what we are facing. As for BCG, by 2033, 19 trillion will be moved and tokenized on-chain . This is a huge number, not only in volume.
Just imagine how many different types of tokens we will have in the future. And the second part is more important. Today, 62% of important. Today, 62% of organizations are experimenting with artificial intelligence, according to McKinsey, but only 10% only 10% are using it in production. This is a huge gap huge gap .
And my question to you is: why? Why are we in the experimental stage? Why aren't we making progress towards a real production task production task ? And to answer more, I think it's worth talking about the vision of how DeFi should be implemented. Three weeks ago, you may have seen a post on Twitter about AI being so cool that it was finding various vulnerabilities that DeFi was no longer a safe place.
We used to face a lot of hacker attacks in the industry, but now people are quite scared. And as you can see scared. And as you can see , Manuel Arroz, the founder of OpenZeppelin, a fairly well-known figure in the security industry in the security industry , wrote this tweet: “ , wrote this tweet: “ Hey, DeFi is dangerous.” And I think, of course, we're in the industry, but imagine what happens when people who manage real capital can verify this. How can we actually solve this and solve the more complex things.
And I have a vision that if we move towards DeFi 3.0, we have two theses. The first is that in the future we will move from a complex smart contract system to a system that uses Ethereum as the settlement layer settlement layer . Previously, for example, if you consider a fairly complex smart smart contract system, there were a lot of things inside, such as internal balance, etc. This less monolithic logic, when we move to P2P, P2P, we simply calculate only one smart contract that supports the line of credit, and only two people get access: the lender and the borrower.
This is quite simple to do through official verification, and after that you have proof that your smart contract has a limited number of users. The second part, which is more interesting part, which is more interesting , is that there are so many assets that it is quite difficult to create a market for them , to consolidate the market. RWA, uh, things related to factoring, etc., etc. factoring, etc.
, etc. This is a very complicated way to build a market on top of them. So, the third part of RWA is not a good design for deposit and redemption. If there are any delays, that's another topic. The second point, if you have settlements, if you have a one-to-one contract, it is very important that both parties conduct due diligence and autonomous management autonomous management .
Otherwise, we will get millions of assets, and no one will be able to really verify what is happening. We currently use risk curators, but I think the future is without curators, where curators, where anyone can use their agents to vet and conduct due diligence on them. And these agents could do deep research, they could respond very quickly to market changes, and that's what we're talking about. And why is autonomy important? So, I think a lot of you are using agents to do research, but if you create autonomy, you will have a much better speed to really follow the market.
You can really scale them. You can really analyze a lot of things, and today I'm going to show you a real use case use case that we created. And the second, third part, she never sleeps. So, I think now you can see people checking Twitter for decision making. Agents work autonomously.
And at Gearbox, we go through the whole complex problem and build the Gearbox and GNC cycle with six six stages. So, we start with discovery, where we do a kind of initial analysis of different options, different assets, different goals. Then we start to dig deeper to do proper due diligence, and the agents are pretty cool, and I'll show you how to go through the whole process, piecing together different pieces of information into one context from Twitter into one context from Twitter , the internet, etc. But because LLMs are not fast, we need another stage, and this is kind of new because LLMs take time. So, we need to have a shortlist of potential candidates, and then we go to propose.
We will try, perhaps We will try, perhaps , to get into this position, and it doesn't always happen because this market changes very, very quickly. So if you think it's profitable for you until the price reaches, I don't know, 1,000, then the price changes until you make that calculation. So, we go back and forth . And then, of course, it's more important to safely execute and continue monitoring. And this is a demonstration of how our artificial intelligence agent works.
So, you can see this in our Agentic app. You can see my two positions, one in Easy USDC and the other in Midas LP and strategies. And we ask the agent what is the best way to place 10 thousand of capital? So I just write a simple query, how to place $10k, and you can see that the agent is constantly reviewing all the data we've collected. The biggest problem we face is that there are not enough good sources of verified data.
So we built our own indexer, and now you see it collects various data, provides graphs, and so on. This is quite interesting information to analyze and build on what I said in the search phase. When we have multiple candidates, the agent now searches the internet, browses Twitter, but it's done so quickly that even you can't walk through the old town. But in this example, I think about 100 documents were analyzed , and the agent detects that some of our pools are not secure enough not secure enough , because there are some incidents with the RLP token. Then I click just to expand, and here you can see a pretty cool way to deposit.
So, after all this, I can make a decision on what to choose. And as you can see, I only spent 5 minutes, but the agent was able to analyze at least 100 documents related to the market. But this is not enough. So this is happening on happening on Gearbox's web servers, and I'm a security fanatic. So next we move on to the MCP server.
You can run this loop in your cloud code, your Codex, or your Codex, or any agent you like. The logic of how we want to achieve security is that when you have a solution, it creates a reference for verification. So you never sign something you don't understand. The person in the cycle is critically important at this stage, and I believe it will be for a long time to come . So you click on this link and you open a special website where you can understand, written but in human language, what you're doing.
If someone can provide your agent with some toxic context, if someone can trick them, you will be the last mile to check if things have gone wrong. And this verification uses common Web 3 security tools. So we don't rely on any server. So you can deploy it to IPFS. This is an unchanging code, so you can be pretty safe in checking what your agent offers you .
But the most interesting thing, I think, about our system is how we can move on and take the next step towards autonomous capital management capital management . Logically speaking, at Gearbox we tested what I had started earlier for 5 years, the model. We know how to calculate the total cost of your credit account. And based on this idea based on this idea , we created the APY bot. An APY bot is a way to create a network within a network and allow some transactions with a certain risk budget.
So, as you can see, this is a window from our application where you have a position and you can let your agent, which runs in the cloud and pulls in various new sources, Twitter, etc., manage your capital. And you set a risk budget because every time agents try to execute a transaction, we check if today's risk budget is sufficient. So, we calculate the total cost before the transaction. It could be 100,000 or more.
And of course you should do that because there are some slippages so you can't do the same thing, but you can really minimize the potential losses. And if we can minimize potential losses by using agent-driven management, that's a pretty cool way for institutional players because it's very easy to hedge. If your losses are limited, maybe 2.0%, you can just buy insurance and it won't be a huge amount. And the third part is if you have failed with your operational security.
As you know, there were many incidents last day. For example, your agent key has been leaked or something has gone wrong. You still stick to that daily risk budget, you can recognize the problem, but you will never have access to the full amount of money. And another part that's quite complicated when you're building an agent that makes decisions based on these news and so on is how you can make it memory-based. It's not enough to just read different sources on Twitter, because as you know, it's very manipulated stuff.
In the past, people tried to manipulate price bands to increase the price and so on. Now you can write a lot of tweets to have a toxic effect on your agent. And the solution that we're working on is that we have a graph of data, and each source has a certain credibility rating to help the system understand whether a statement is somewhat credible or whether it's completely false. And in this case, maybe someone was reading the news when people added some comments to the quote, and the quote changed behavior. With these things, at least, there is a defense of how to get an agent to adhere to this way of thinking.
And another part of using graph memory that's pretty cool is that it's very extensible. So, if your institution has its own due diligence department, you have your own methodology, you have years of evaluating different assets. And this data easily connects to this graph and helps the agent make better decisions. Previously, when we talked about money management money management , we tried to beat a machine learning system that should win in all cases and provide high efficiency. But now, I think we'll change a little bit in this LLM, we can provide some kind of manifesto of what we want to achieve, what areas we think will develop further, and then this system can follow your manifesto and try to do its best.
Why is this, uh, different, because two different institutions created the same system but set different goals to achieve them. And the second part, related to the graph, explains why agents are much, much better than all existing risk curators and humans. Why? Because assets today are very complex. We started in 2021, and I clearly remember a time when token symbols were only three letters.
It was something like four, USDC. We currently have tokens like Woo, Moo, Boo, Base, Wrapped, etc. The difficulty is approaching. And to find all the connections between tokens, some of them use the provisioning in Gearbox, then it's connected to Morpho, to 3F and several protocols. Imagine that one asset can be interconnected with 100 companies, and you cannot easily understand such interactions.
You can't do these things easily. If you build this depth-to-depth graph correctly this depth-to-depth graph correctly , you can create different simulations and teach your agent how to behave. So you won't even be able to find, and this is critical, whether any asset might be affected. You may remember this extreme financial hack where Elixir was based on them and then some risk curator banned them, etc., etc.
It's like a domino effect. And of course, it's pretty cool to have that depth, and then you can set up an emergency exit. If an asset fails, you can immediately find the correlation using this deep research and save your money. I think this is the best thing we can build on. And as I promised, here's how you can create an even better system if you're not interested in using leverage but want to have better asset management for yourself personally.
So the basic idea is to create a structured product based on the Gearbox stack to have some advantages over your management. For example, you have a fairly stable source of income. Let's imagine you have 10 thousand USDC tokens. Then you divide it into two categories. 90 it into two categories.
90 % of this amount you invest in the pool, and since they are in the form of, yes, USDC, they generate income for you . This is your passive side. You then create a credit account managed by an agent and can cycle through different positions, investing in different tokens and opportunities. This agent can lend money, but now it is quite simple that if the agent does not provide the same performance as USD tokens, it will be liquidated. We use the word "liquidate" as a loan protocol, but here it works as a stop loss.
So, literally you have 90% of your capital saved, then your agents borrow from your pocket and try to beat it. And if this agent does not achieve the target performance, such a position has the status of a stop-loss on the chain, which is also well- tested and proven in battles over the past five years. And in this case, you should not rely solely on the agent's stochastic decisions. You have a pretty robust system on the chain that creates the driving force and really strict rules that you can't underperform. You must compete and surpass ours.
And so, this is a very good way you can launch it tomorrow. For example, I personally want to do this because I don't have enough time to search Twitter and everything else. And then he will automatically find the best way to outperform the competition. As you know, the best way Google will teach robots to play Go. Robots play with themselves to find the best strategy and learn from each other.
We can do the same thing in finance, and this is one way it should be done. So, we think we'll launch an experimental way of our Gearbox agent. If you have...Oh, I think I'm on time. If you have any questions, maybe we have half a minute.
Thank you very much. Thank you for your attention.
Automatic transcript — names and jargon may be misspelled.